In order to solve the real-time synthesis problem in a multi-sample text-to-speech system, the strategy of improving the structure of speech database is discussed and a method of calculating the similarities between two speech units is put forth in this paper. On the basis of these, a mapping address algorithm to locate the address of speech unit is brought up. The experimental results show that mapping address algorithm and reconstruction of speech database improve the real-time performance of text-to-speech system.